On October 9, 2026, Asana published results from experiments that used GPT‑6 Astra in Codex to optimize a browser-agent workflow running on GPT‑6.1 Sol. The company said the optimized workflow ran an estimated 76x cheaper and about 5x faster than its original production setup on Model B.
Study design and investigation
Asana described a 144-run study that compared GPT‑6.1 Sol with three other frontier models (Models A, B and C). The study simulated a representative StackAI task: collecting six fields for each of 32 books from a public demo catalog. Engineers directed GPT‑6 Astra to map the codebase, identify inefficiencies in how the agent assembled model requests, and propose fixes.
GPT‑6 Astra’s analysis reportedly found the agent cached fixed instructions and tool definitions but repeatedly resent growing page text and screenshots, driving up input costs. The team tested variables including history budgets (120,000 and 480,000 characters), several caching and screenshot policies, and batching of screenshot removals. The best policy allowed screenshots to accumulate to 20 before pruning.
Results, timings and rollout
Asana reported the optimized workflow on GPT‑6.1 Sol averaged $0.47 in estimated model costs and about four minutes per run. Compared with the original Model B setup, Asana said this was 76x cheaper and 5x faster. For Model B specifically, the optimization reduced an estimated cost from at least $36.21 to $1.24 per run (a 29x reduction); the GPT‑6.1 Sol workflow was 2.6x cheaper than that $1.24 figure.
The company also reported that, on GPT‑6.1 Sol with the larger history budget, cost fell 4x from $1.97 to $0.47 per run because 89% of input came from cache priced at about 5% of uncached input. Runs that produced an answer increased from 3 of 18 with the smaller history budget to all 18 with the larger budget. Asana said every run in the optimized workflow returned the correct answer in the evaluated samples.
Asana stated the changes have been released to StackAI and that it is building tools to make these experiments easier to repeat and to incorporate cost, runtime and answer-quality comparisons into platform evaluations.
Original source: OpenAI News